Stabilizing fisheye video from a light-mounted camera

We present in this paper a method to stabilize video acquired by a fisheye camera that has been mounted underneath a cable-suspended traffic light. With cable suspension, the light, and hence the camera, tend to swing in the wind implying that stationary structures appear to move in the acquired video. To make use of the video for higher level computer vision functions, such as the detection and tracking of moving vehicles, we must account for this motion. In the proposed approach, we first select a set of salient points on fixed structures in a reference image. We then track the motion of these points in subsequent digital images using the Lucas-Kanade algorithm. Based upon the measured movement of the fixed points, we compute a mapping between the fisheye coordinate systems of the subsequent images and that of the reference image. To stabilize the video, we use these mappings to resample the fisheye images so that they are aligned with the reference fisheye image. We demonstrate the results of our approach on real data.